Machine Vision Image Compression via Feature Map Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image compression technologies for machine vision tasks are inefficient in compressing videos due to ineffective redundancy removal and high data redundancy, particularly when extending existing video codec technology to videos for machine-specific tasks.
Innovation Solution
An image compression method that determines a prediction mode for frames, generates prediction and residual images, corrects the reconstructed frame based on feature differences and bit rate, and encodes the corrected frame, allowing for iterative deformation of frames to maintain compression efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of substance
If existing video codec technology is used for machine vision tasks, then compression is achieved, but redundancy removal is ineffective and data redundancy remains high
Solution Approach 1:
The patent extracts and removes redundant information from video data by computing feature maps through convolution operations, identifying and eliminating temporally and spatially redundant information while preserving essential features for machine vision tasks
Solution Approach 2:
The patent segments video data into frames and further into feature maps through convolutional operations, dividing the processing into manageable units that can be independently analyzed and compressed based on their redundancy characteristics
2Loss of substance
If image compression is performed to maximize compression ratio, then bitstream size is reduced, but machine vision task performance may deteriorate
Solution Approach 1:
The patent applies different processing quality to different parts of the image by computing feature maps that highlight task-relevant regions, preserving high quality information where needed while allowing greater compression in redundant regions
Solution Approach 2:
The patent transforms images into feature map representations through convolution operations, changing the parameter space from raw pixel values to task-relevant feature activations that can be more efficiently compressed
3Device complexity
If frames are compressed independently, then processing is simplified, but compression efficiency decreases due to ignoring temporal redundancy
Solution Approach 1:
The patent performs preliminary processing by computing feature maps and identifying prediction modes before actual compression, preparing the data in advance to enable more efficient compression by exploiting temporal relationships
Data Source
AI summary
Disclosed herein are an image compression method and apparatus for machine vision. The image compression method includes determining a prediction mode for frames of an input image, generating a prediction frame and a residual image using an input frame, generating a reconstructed frame by adding the prediction frame to the residual image, extracting respective features of the input frame and the reconstructed frame, correcting the reconstructed frame based on a difference value between the extracted features and a bit rate of the residual image, and encoding the corrected frame.